đ§ Changing the plan, not the model, lifted a robot world model's success from 22% to 85%
Researcher Haddy Alchaer ran 600 experiments on LeWM, a JEPA world model from a team that includes Yann LeCun. As published, it succeeded on 22% of simulated tasks where a robot must push a T-shaped block to a goal 75 steps away. A new subgoal every 25 steps raised that to 80%, and a memory of past demonstrations with a cost on every jump between them reached 85%. On paths the model was trained for, its own predictions scored 78%, and the simulator's real physics scored 77%. No weights were retrained.
â The useful fix was a shorter plan, not a smarter predictor.
#AI #Robotics
Researcher Haddy Alchaer ran 600 experiments on LeWM, a JEPA world model from a team that includes Yann LeCun. As published, it succeeded on 22% of simulated tasks where a robot must push a T-shaped block to a goal 75 steps away. A new subgoal every 25 steps raised that to 80%, and a memory of past demonstrations with a cost on every jump between them reached 85%. On paths the model was trained for, its own predictions scored 78%, and the simulator's real physics scored 77%. No weights were retrained.
â The useful fix was a shorter plan, not a smarter predictor.
#AI #Robotics
